Stochastic Dominance and Cumulative Prospect

2007-02
15/1/07
10:13
Página 1
Documentos
de Trabajo
2
Documentos
de Trabajo
2
2007
Manel Baucells Alibés
Franz H. Heukamp
Stochastic Dominance
and Cumulative
Prospect Theory
Gran Vía, 12
48001 Bilbao
Tel.: 94 487 52 52
Fax: 94 424 46 21
Paseo de Recoletos, 10
28001 Madrid
Tel.: 91 374 54 00
Fax: 91 374 85 22
[email protected]
www.fbbva.es
Stochastic Dominance and Cumulative
Prospect Theory
Manel Baucells Alibés
Franz H. Heukamp
U N I V E R S I T Y
O F
N A V A R R A
Abstract
Resumen
We generalize and extend the second order
stochastic dominance condition available for
Expected Utility to Cumulative Prospect Theory.
The new definitions include, among others,
preferences represented by S-shaped value and
inverse S-shaped probability weighting functions.
The stochastic dominance conditions supply a
framework to test different features of Cumulative
Prospect Theory. In the experimental part of the
working paper we offer a test of several joint
hypotheses on the value function and the
probability weighting function. Assuming
empirically relevant weighting functions, we can
reject the inverse S-shaped value function recently
advocated by Levy and Levy (2002a), in favor of
the S-shaped form. In addition, we find generally
supporting evidence for loss aversion. Violations of
loss aversion can be linked to subjects using the
overall probability of winning as heuristic.
Este documento de trabajo generaliza y extiende
las condiciones de dominancia estocástica de
segundo orden disponibles para utilidad esperada a
la teoría de prospectiva cumulativa. La nuevas
definiciones incluyen, entre otros, las preferencias
representadas por una función de valor en forma de
S y una función de probabilidad ponderada en
forma de S invertida. Las condiciones de
dominancia estocástica permiten diseñar test para
diferenciar las distintas partes de la teoría de
prospectiva cumulativa. En la parte experimental
del documento de trabajo se examinan varias
hipótesis conjuntas sobre al función de valor y la
función de probabilidad ponderada. Asumiendo
una forma empíricamente relevante de la función
de probabilidad ponderada, podemos rechazar la
hipótesis de una función de valor en forma de S
invertida, defendida recientemente por Levy y Levy
(2002a), a favor de la hipótesis de una función de
valor en forma de S. Además, encontramos
evidencia a favor de aversión a la pérdida.
Violaciones de aversión a la pérdida se pueden
atribuir a que algunos sujetos usan como regla
heurística la probabilidad de obtener ganancias.
Key words
Palabras clave
Second order stochastic dominance, cumulative
prospect theory, value function, probability
weighting function.
Dominancia estocástica de segundo orden, teoría
de prospectiva cumulativa, función de valor,
función de probabilidad ponderada.
Al publicar el presente documento de trabajo, la Fundación BBVA no asume responsabilidad alguna sobre su contenido ni sobre la inclusión en el mismo de
documentos o información complementaria facilitada por los autores.
The BBVA Foundation’s decision to publish this working paper does not imply any responsibility for its content, or for the inclusion therein of any supplementary documents or
information facilitated by the authors.
La serie Documentos de Trabajo tiene como objetivo la rápida difusión de los
resultados del trabajo de investigación entre los especialistas de esa área, para
promover así el intercambio de ideas y el debate académico. Cualquier comentario sobre sus contenidos será bien recibido y debe hacerse llegar directamente a
los autores, cuyos datos de contacto aparecen en la Nota sobre los autores.
The Working Papers series is intended to disseminate research findings rapidly among
specialists in the field concerned, in order to encourage the exchange of ideas and academic debate. Comments on this paper would be welcome and should be sent direct to the
authors at the addresses provided in the About the authors section.
Todos los documentos de trabajo están disponibles, de forma gratuita y en formato PDF, en la web de la Fundación BBVA. Si desea una copia impresa, puede
solicitarla a través de [email protected].
All working papers can be downloaded free of charge in pdf format from the BBVA
Foundation website. Print copies can be ordered from [email protected].
La serie Documentos de Trabajo, así como información sobre otras publicaciones de la
Fundación BBVA, pueden consultarse en
http://www.fbbva.es
The Working Papers series, as well as information on other BBVA Foundation publications,
can be found at
http://www.fbbva.es
Stochastic dominance and cumulative prospect theory
© Manel Baucells Alibés and Franz H. Heukamp, 2007
© de esta edición / of this edition: Fundación BBVA, 2007
EDITA
/ PUBLISHED
BY
Fundación BBVA, 2007
Plaza de San Nicolás, 4. 48005 Bilbao
/ LEGAL DEPOSIT NO.: M-3.428-2007
/ PRINTED BY: Rógar, S. A.
DEPÓSITO LEGAL
IMPRIME
Impreso en España – Printed in Spain
La serie Documentos de Trabajo de la Fundación BBVA está elaborada con papel 100% reciclado,
fabricado a partir de fibras celulósicas recuperadas (papel usado) y no de celulosa virgen, cumpliendo los estándares medioambientales exigidos por la actual legislación.
The Working Papers series of the BBVA Foundation is produced with 100%recycled paper made from recovered cellulose fibre (used paper) rather than virgin cellulose, in conformity with the environmental standards required by current legislation.
El proceso de producción de este papel se ha realizado conforme a las normas y disposiciones
medioambientales europeas y ha merecido los distintivos Nordic Swan y Ángel Azul.
The paper production process complies with European environmental laws and regulations, and has both
Nordic Swan and Blue Angel accreditation.
C O N T E N T S
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
2. CPT Stochastic Dominance Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1. Accounting for the value function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2. Accounting for the probability weighting function . . . . . . . . . . . . . . 11
2.3. Accounting for loss aversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3. Experimental Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1. Choice of c and d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2. Shape of the value function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3. Shape of the pwf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4. Loss aversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5. CPT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
21
22
24
25
27
4. Conclusions and Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Appendix: Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
About the Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3
1. Introduction
STOCHASTIC Dominance (SD) relations offer an efficient way to
compare pairs of prospects. This has been recognized ever since their
introduction to economics by Rothschild and Stiglitz (1970). SD
conditions are applied in finance, decision analysis, economic modeling,
and axiomatic modeling (Levy, 1992). Well known specifications of SD are
First-Order SD (FSD) and Second-Order SD (S-SD) which have been
presented in the context of Expected Utility (EU). The interpretation of
SD conditions is often useful by itself. FSD can be seen as a stochastic
version of preferring more money to less, and S-SD as a characterization of
riskaversion. Preferences satisfying these SD conditions have a
representation that suits this interpretation: FSD is associated with
non-decreasing utility, and S-SD with concave utility. SD conditions have
other appealing features. Knowing certain qualitative features of the utility
function of a decision maker, one can use the corresponding SD condition
to eliminate dominated alternatives, which can be useful for management
practitioners. More importantly, SD conditions can be used to construct
pairs of prospects in which one prospect dominates the other. From the
preference of a decision maker between these prospects one can deduce
qualitative conditions of her preferences (e.g., risk aversion) and conclude
qualitative properties of their representation (e.g., concavity of the value
function). SD can thus be a potentially powerful framework for testing
qualitative properties of choice models. It is this application that we focus
on in this working paper.
EU does not accurately predict empirically observed preferences
(Camerer, 1995; Wu, Zhang, and Gonzalez, 2004) while, in contrast,
Cumulative Prospect Theory (CPT) (Tversky and Kahneman, 1992) is
quite successful at it (see Abdellaoui [2000] and Abdellaoui, Vossmann,
and Weber [In Press] for recent tests). Therefore, SD conditions related to
CPT are likely to be of practical interest. They can suggest experimental
designs that can isolate certain features of CPT without having to estimate
all the elements of that theory. CPT proposes that subjects encode
outcomes in terms of gains and losses. Furthermore, CPT replaces the
traditional utility function by a value function defined over variations of
wealth with respect to some reference point, and includes a probability
5
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
weighting function (pwf) that reflects the subjective probability distortion
shown by most individuals.
We begin by offering a general SD result for the EU model, which
accommodates several curvature forms for the utility function, namely,
S-shaped, inverse S-shaped, concave, and convex. Next, we generalize the
previous result to the CPT model by incorporating the pwf. For the value
function, this second result encompasses the previous four curvature
combinations. For the pwf, we go beyond the case of a constant curvature,
investigated in previous generalizations, and consider a broader class that
includes the inverse S-shaped form. In addition, we examine SD conditions
that capture the remaining important aspect of CPT, namely, loss aversion.
Loss aversion plays a central role in behavioral decision research (Benartzi
and Thaler, 1995; Langer and Weber, 2001); in CPT it is expressed
mathematically as a steeper value function for losses than for gains,
capturing the psychological intuition that losses loom larger than gains.
We conclude the theoretical part of the working paper by presenting an
SD condition that incorporates loss aversion, an S-shaped value function,
and an inverse S-shaped pwf, i.e., Prospect Stochastic Dominance in the sense
of the specification of CPT put forward by Tversky and Kahneman (1992).
Stochastic Dominance conditions are necessary and sufficient to
characterize combinations of classes of value functions and pwfs. They can
be used to test joint hypotheses on the curvature of the value function and
the pwf. This can be done in three ways, the application of which is
presented in the experimental part of this working paper. First, if one
assumes that the pwf is inverse S-shaped, then the corresponding SD
conditions serve the purpose of testing hypotheses about the curvature of
the value function and/or about loss aversion. Second, if one assumes that
the specifications of CPT for the curvature of the value function hold, then
the corresponding SD conditions allow one to test hypotheses on the
shape of the pwf. Third, if one assumes that CPT holds under established
qualitative specifications for the value function, including loss aversion,
and the pwf, then a violation of the corresponding SD condition implies a
violation of the CPT model. This suggests ways to compare CPT with
alternative heuristics models of decision that have been proposed in the
literature. The experiments reported in this working paper involved 277
subjects, consisting of undergraduates, MBA students and executives.
6
2. CPT Stochastic
Dominance
Conditions
THROUGHOUT the working paper we will be comparing pairs of
prospects F and G, having cumulative distributions F and G. For
simplicity, we assume that the outcomes of F and G are contained in an
interval [a, b], for some a < 0 and b > 0, and that F and G are continuous
except for finitely many points. Subjects are assumed to abide by the rankand sign-dependent framework of CPT. This is, their preferences are
represented by (1) a non-decreasing value function v (x) defined over a
monetary gain or loss (change in wealth with respect to some reference
point), with v(0) = 0; and (2) a pair of non-decreasing probability
weighting functions (pwf) w− (p) and w+ (p) that transform the objective
probabilities into decision-weights, and having w(0) = 0 and w(1) = 1 (w is
short for “both w+ and w− ”). We also assume that v is continuous and,
except for finitely many points, differentiable. w is assumed to be
continuous, except for finitely many points.
Cumulative Prospect Theory (CPT) (Tversky and Kahneman, 1992)
has become widely accepted as a descriptive theory of choice under risk.
Empirical specifications of the model indicate that v is S-shaped. Formally,
we define VP , the family of Prospect value functions, as those v(x) that are
convex for x < 0 and concave for x ≥ 0. In order to contrast this
specification with other alternatives, we define VP ∗ as the family of Inverse
Prospect value functions containing those v(x) that are concave for x < 0,
and convex for x ≥ 0. Similarly, VConcave (VConvex ) is the family of value
functions that are concave (convex) for all x.
The empirically observed probability distortion agrees with a pwf
that is “shallow in the open interval and changes abruptly near the
end-points where w(0) = 0 and w(1) = 1” (Tversky and Kahneman,
1992: 282), more specifically, an inverse S-shaped pwf that is concave first,
and then convex, has found broad empirical support (see Tversky and
Kahneman, 1992; Tversky and Wakker, 1995; Gonzalez and
Wu, 1999; Abdellaoui, 2000).
7
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
2.1.
Accounting for the value function
In this section, we assume the decision maker maximizes the expectation
of the value function, and she does not distort probabilities (i.e., w(p) is
not part of this model). Depending on the interpretation of x (relative or
absolute) we encompass either a reference dependent model or EU. In
any such model, the choice between F and G is determined by VF − VG ,
the difference between the evaluation of each prospect, where
Rb
VF = a v(x)dF (x) is the expectation of v under F . If F and G are
continuous, then integration by parts produces:
Z
∆[a,b] ≡ VF − VG =
b
[G(x) − F (x)]v 0 (x)dx
(2.1)
a
∆[a,b] can be seen as the inner product of G(x) − F (x) and v 0 (x). The
relative magnitude of ∆[a,b] can be seen in the graph of the cumulative
distributions of F and G (e.g. graphic 2.1) by first deforming the
horizontal payoff axis in proportion to v 0 , and then simply calculating the
area between G and F . In this process, segments where the slope of v(x) is
high are stretched more relative to segments where it is small. No
deformation (v 0 = 1), of course, corresponds to the risk-neutral evaluation
Rb
a [G(x) − F (x)]dx = EF − EG . In any case, the integration in (2.1) never
decreases in intervals where F dominates, G(x) > F (x), and it never
increases in intervals where G dominates.
In the left graph of graphic 2.1, the “+” and “−” signs correspond to
the signs of G(x) − F (x), respectively.
Consider some interval [a0 , a1 ] in which (i) the value function is
convex, and (ii) any “−” area is followed by some “+” area of equal or larger
size. By convexity, the horizontal stretching of the “+” area will be larger
than the horizontal stretching of the “−” area, and the evaluation of (2.1)
on that interval, called ∆[a0 ,a1 ] , will be positive. Similarly, consider some
interval [b0 , b1 ] in which (i) the value function is concave, and (ii) any “−”
area is preceded by some “+” area of equal or larger size. By concavity, the
horizontal stretching of the “+” area will be larger than the stretching of
the “−” area, and ∆[b0 ,b1 ] will be positive.
For a prospect value function, the function is convex in [a, 0] and
concave in [0, b]. Hence, the required cancelation occurs in graphic 2.1
(left), ensuring that F is preferred over G. Conversely, if v is inverse
S-shaped, then the convex and concave intervals are [0, b] and [a, 0],
respectively. The area cancelation now finds that G is preferred over F .
8
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
GRAPHIC 2.1:
Left: Graphic illustration of stochastic dominance for Task V. Right: h(x) for Task V
For the ensuing definitions, we introduce the function
Z x
h(x) ≡
[G(y) − F (y)]dy
(2.2)
a
Clearly, h is continuous and h0 (x) = G(x) − F (x). Moreover,
h(a) = 0 and h(b) = EF − EG , so that h(x) can be viewed as the
“advantage” in expected value of F over G up to x. The right graph of
graphic 2.1 (right) shows an example of this function.
The value of h(x1 ) − h(x0 ) measures the area contained between G
and F in the interval [x0 , x1 ]. The function h is very convenient to express
stochastic dominance conditions. For instance, h0 (x) ≥ 0 for all x is
equivalent to G(x) ≥ F (x), which is FSD. Similarly, h(x) ≥ 0 for all x is the
second-order stochastic dominance condition. In the context of the
previous discussion, the condition that a “+” area is canceled by some
subsequent “−” area on [a0 , a1 ] is easily expressed as h(x) ≤ h(a1 ) for all x,
a0 ≤ x < a1 . Similarly, the condition that a “−” area is canceled by some
preceding “+” area on [b0 , b1 ] is easily expressed as h(x) ≥ h(b0 ) for all x,
b 0 ≤ x < b1 .
A general SD result can now be presented considering the interplay
between the function h and the curvature of v.
Proposition 1. Let A = [a0 , a1 ] and B = [b0 , b1 ] be intervals such that
9
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
A ∪ B = [a, b]. Given prospects F and G, let h be defined as in (2.2). Then,
h(x) ≤ h(a1 ) for a0 ≤ x < a1 , and
(2.3)
h(x) ≥ h(b0 ) for b0 ≤ x < b1
(2.4)
hold if and only if VF ≥ VG for all v that are convex in A and concave in B.
Notice that the intervals A and B don’t have to be disjoint, in which
case v is linear in their intersection. We present four specialized versions
of proposition 1 tailored to the four relevant curvature combinations of
reference-dependent value functions. The corresponding integral
conditions will be denoted by P -SD (prospect), P ∗ -SD (inverse prospect),
S-SD (second order), and S ∗ -SD (inverse second order) stochastic
dominance, respectively1 .
Definition 2. Given prospects F and G, let h be defined as in (2.2). F dominates
G according to P -SD, denoted by F P -SD G, if and only if h(x) ≤ h(0)
for a ≤ x ≤ 0, and h(x) ≥ h(0) for 0 ≤ x ≤ b. Similarly, we define
F P ∗ -SD G
iff
F S-SD G
F S ∗ -SD G
iff
iff
h(x) ≥ h(a) for a ≤ x ≤ 0, and
h(x) ≤ h(b) for 0 ≤ x ≤ b;
h(x) ≥ h(a) for a ≤ x ≤ b; and
h(x) ≤ h(b) for a ≤ x ≤ b.
Proposition 3. F P -SD G if and only if VF ≥ VG for all v ∈ VP . Similarly,
F P ∗ -SD G
F S-SD G
F S ∗ -SD G
iff
iff
iff
VF ≥ VG for all v ∈ VP ∗ ;
VF ≥ VG for all v ∈ VConcave ; and
VF ≥ VG for all v ∈ VConvex .
These four particular results are available in Levy and Levy (2002b),
albeit not in this unified treatment. To illustrate the proposition, notice
that F and G in graphic 2.1 satisfy F P -SD G and G P ∗ -SD F . We also
remark that if any of these conditions apply, then EF ≥ EG . To see this,
notice that a is equal to either a0 or b0 , and b is equal to either a1 or b1 . If
follows from (2.3) and (2.4) that EF − EG = h(b) ≥ h(a) = 0.
1. Combination of letters such as P , P ∗ , S, S ∗ , W and L will denote different SD conditions.
Although this type of notation might seem cumbersome at the beginning, in the course of the
working paper its usefulness will become apparent - please bear with us.
10
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
2.2.
Accounting for the probability weighting function
In this section (and the rest of the working paper), we assume that the
decision maker follows CPT, i.e., she has preferences represented by a
value function and a probability weighting function. The introduction of a
nonlinear and sign-dependent pwf transforms VF − VG in (2.1) into
∆w
[a,b]
Z
b
Z
=
0
[w− (G(x)) − w− (F (x))]v 0 (x)dx+
a
[w+ (1 − F (x)) − w+ (1 − G(x))]v 0 (x)dx
(2.5)
0
In the discrete case, we first label the outcomes as
a < x1 < ...xk−1 < xk = 0 < xk+1 < ... < xn < b,2 and let
∆
w
=
+
k−1
X
[w− (G(xi )) − w− (F (xi ))][v(xi+1 ) − v(xi )]
i=1
n−1
X
[w+ (1 − F (xi )) − w+ (1 − G(xi ))][v(xi+1 ) − v(xi )]
i=k
(2.6)
with the understanding that if k = 1 (k = n) then the summation over
negative (positive) outcomes disappears.
Considering the evaluation of ∆w
[a,b] according to (2.5), we observe
that w adds a non-linear transformation of the vertical axis in proportion
to w0 (p) (Quiggin, 1993: 150). For concreteness, let’s consider an inverse
S-shaped pwf. Referring to graphic 2.1 (left) the nonlinearity of the pwf
stretches the lower and upper ends of the vertical probability axis, relative to
the middle range of cumulative probabilities. Accordingly, the gap
between G(x) and F (x) in zones near the probability extremes is
magnified. As a consequence, CPT remains ambiguous in the case of
prospects F and G in graphic 2.1 (left). While F is favored by the
stretching of the horizontal axis near the origin enlarging the “+” areas, G
is favored by the stretching of the vertical axis for low values of p in the
losses domain and for high values of p in the gains domain, enlarging the
“−” areas. Thus, according to (2.5), an S-shaped value function is still
compatible with VG > VF . In the following we argue that second order SD
conditions require that the value function and the pwf have conjugate
2. Notice that to have xk = 0 for some k may require setting pk = 0.
11
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
GRAPHIC 2.2:
Schematic depiction of the Wcd class of pwfs. Notice that the
Wcd class grows as c increases and d decreases
curvatures. By conjugate we refer to curvature combinations that ensure the
monotonicity of [w− (G(x)) − w− (F (x))]v 0 (x)/[G(x) − F (x)] or
[w+ (1 − F (x)) − w+ (1 − G(x))]v 0 (x)/[G(x) − F (x)]. In the losses domain,
a convex v and a convex w− are conjugate; or a concave v and concave w−
are also conjugate. In the gains domain, a concave v and a convex w+ are
conjugate; or a convex v and a concave w+ are conjugate too.
Assuming a constant curvature for the pwf is the simplest possibility,
but this would limit the potential for application of our results. With the
goal of finding stochastic dominance conditions that can encompass an
inverse S-shaped pwf, we consider pwfs that are concave in the range [0, d)
and convex in the range (c, 1], for given values of d and c in [0, 1]. We
denote this class by Wcd . If c < d, then the segment between c and d is
necessarily linear; and if 0 < c ≤ d < 1, then the pwf is inverse S-shaped
and continuous in (0, 1) (see graphic 2.2). Most of the commonly
employed parametric families of pwfs (Prelec, 1998; Tversky and
Kahneman, 1992; Lattimore, Baker, and Witte, 1992) fall within the Wcd
class, with c = d being the inflection point of these inverse S-shaped
functions. The family W01 , c = 0 and d = 1, corresponds to pwfs that are
linear in (0, 1) and possibly discontinuous at either 0 or 1. Finally, if c > d,
then w is unrestricted between d and c. Here, we have used c (or d) to
denote both c− and c+ (or both d− and d+ ), which will apply to w− and
w+ , respectively.
12
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
We are interested in the intervals of the payoff line in which the
curvatures of v and w are conjugate. To describe their boundaries, we
define four characteristic points such as the point x−
c , which corresponds
to the algebraically smallest negative payoff for which both F and G take
values greater than c− . Analogously, x−
d is the greatest value in the
negative domain for which both F and G take smaller values than d− ; x+
c
the greatest positive payoff for which F and G are smaller than 1 − c+ ; and
x+
d the smallest positive payoff for which F and G take values greater than
1 − d+ . Formally,
Definition 4. If the pwf is continuous at the corresponding points c− , d− , d+ , and
d+ , then let
x−
= inf{x ≤ 0 : F (x) ≥ c− , G(x) ≥ c− },
c
x−
d
x+
c
x+
d
−
(2.7)
−
= sup{x ≤ 0 : F (x) ≤ d , G(x) ≤ d },
(2.8)
+
+
+
+
= sup{x ≥ 0 : F (x) ≤ 1 − c , G(x) ≤ 1 − c }, and
(2.9)
= inf{x ≥ 0 : F (x) ≥ 1 − d , G(x) ≥ 1 − d }, respectively.
(2.10)
If the pwf is not continuous at some of these points, then replace the
corresponding weak inequalities (e.g., F (x), G(x) ≥ c− ) with strict inequalities
(e.g., F (x), G(x) > c− ). If the sets over which the infimum and the supremum are
−
+
+
taken are empty, then set x−
c = 0, xd = a, xc = 0, and xd = b, respectively.
The definition ensures that the curvatures of w− , w+ , and v are
conjugate:
−
Proposition 5. If x−
c < x < 0 and x ∈ A (a ≤ x < xd and x ∈ B) then
−
−
0
[w (G(x)) − w (F (x))]v (x)/[G(x) − F (x)] is non-decreasing
+
(non-increasing). If 0 ≤ x < x+
c and x ∈ B (xd < x < b and x ∈ A), then
+
+
0
[w (G(x)) − w (F (x))]v (x)/[G(x) − F (x)] is non-increasing
(non-decreasing).
To illustrate the intervals, let c = 0.1 and d = 0.8, and consider the
−
prospects in graphic 2.1 (left). We have that x−
c = −3000, xd = 0,
+
−
+
x+
c = 3000, and xd = 0. In this example, xd and xd are both zero because
F and G never reach a cumulative probability of d = 0.8 in the negative
domain; and it exceeds 1 − d = 0.2 in the positive domain.
For the results that follow, we are given prospects F and G, and
−
+
+
values of c− , d− , c+ , d+ . We then calculate x−
c , xd , xc , and xd according
to definition 4, and h(x) as in (2.2). We generalize proposition 1 by
13
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
ensuring first order stochastic dominance in the zones lacking conjugate
curvature. The result is rather technical: non-interested readers can skip
this statement and move to the particularization that follows.
Proposition 6. Let A = [a0 , a1 ] and B = [b0 , b1 ] be intervals such that
A ∪ B = [a, b], and let â0 = max{a0 , 0} and b̂1 = min{0, b1 }. Then, conditions
(2.3) and (2.4) from proposition 1, together with
+
if x ∈ A, then G(x) ≥ F (x) for a0 ≤ x < x−
c and â0 ≤ x < xd ,
x−
d
x+
c
if x ∈ B, then G(x) ≥ F (x) for
≤ x < b̂1 and
≤ x < b1 ,
if a0 < 0 < a1 , then h(x) ≤ h(0) for a0 ≤ x < 0, and
if b0 < 0 < b1 , then h(x) ≥ h(0) for 0 ≤ x < b1
(2.11)
(2.12)
(2.13)
(2.14)
−
hold if and only if VF ≥ VG for all v convex in A and concave in B, w− ∈ Wcd− ,
+
and w+ ∈ Wcd+ .
Based on this general result, we present the corresponding
extension of proposition 3, now incorporating the non-linear pwf. In
terms of notation, we add a “W ” to the names of the SD conditions.
Definition 7. F dominates G according to P W -SD, denoted by F P W -SD G, if
and only if F dominates G according to P -SD, and G(x) ≥ F (x) for
+
a ≤ x < x−
c and xc ≤ x < b. Similarly, we define
+
∗
F P W -SD G iff F P ∗ -SD G, and G(x) ≥ F (x) for x−
d ≤ x < xd ;
−
F SW -SD G iff F S-SD G, G(x) ≥ F (x) for xd ≤ x < 0 and x+
c ≤ x < b, and
h(x) ≥ h(0) for 0 ≤ x < b; and
+
F S ∗ W -SD G iff F S ∗ -SD G, G(x) ≥ F (x) for a ≤ x < x−
c and 0 ≤ x < xd ,
and h(x) ≤ h(0) for a ≤ x < 0.
Preferences that agree with these dominance conditions have the
following representation.
−
Proposition 8. F P W -SD G if and only if VF ≥ VG for all v ∈ VP , w− ∈ Wcd−
+
and w+ ∈ Wcd+ . Similarly,
−
F P ∗ W -SD G iff VF ≥ VG for all v ∈ VP ∗ , w− ∈ Wcd− and
+
w+ ∈ Wcd+ ;
−
F SW -SD G
iff VF ≥ VG for all v ∈ VConcave , w− ∈ Wcd− and
+
w+ ∈ Wcd+ ; and
−
F S ∗ W -SD G iff VF ≥ VG for all v ∈ VConvex , w− ∈ Wcd− and
+
w+ ∈ Wcd+ .
An example for F P W -SD G, c = 1/6, is displayed in graphic 2.3.
The fact that FSD has to be imposed in the absence of conjugate
14
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
GRAPHIC 2.3:
Cumulative distributions for Task III. To the left and the right of the cumulative graph we
indicate the curvature of the pwf, according to the Wcd class
curvatures is in itself very revealing: If no restriction is placed on the pwf,
then all four SD conditions reduce to FSD. In other words, knowledge of v,
together with an unrestricted pwf, is insufficient to predict preferences between
stochastically undominated prospects. Conversely, a pwf with certain constant
curvatures allows us to extend with minor modifications the known forms
of SD to the CPT framework.
For example, if both w− and w+ are convex (c− = c+ = 0), then
P -SD and P W -SD coincide (Levy and Wiener [1998]). To extend S-SD to
SW -SD requires w− to be concave (d− = 1) and w+ to be convex
(c+ = 0), together with h(0) ≤ h(x) for all x, 0 ≤ x < b. In general, by
decreasing c or increasing d we make the Wcd class narrower, but impose
FSD on a smaller range, and hence increase the scope of application of the
different SD conditions. However, if the class Wcd is too narrow, then it
might not contain the desired functions. Ultimately, the choice of c and d
is the product of resolving this trade-off. From an experimental point of
view this trade-off will be addressed in section 3.1.
15
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
2.3.
Accounting for loss aversion
Loss aversion, an important feature of CPT with strong empirical support
(Abdellaoui, Bleichrodt, and Paraschiv, 2004), has not played a role in our
definitions so far. To incorporate this third feature of CPT, we define the
class of value functions possessing loss aversion as follows. The value
function v is in VL if and only if
v 0 (−x) ≥ v 0 (x) for all x > 0 where v 0 (x) and v 0 (−x) exists.
(L)
Condition (L) was characterized through preference conditions by
Wakker and Tversky (1993) and is essentially equivalent to
v(y) − v(x) ≤ v(−x) − v(−y) for all y ≤ x ≤ 0 (Bowman, Minehart, and
Rabin [1999, Assumption A2]). (L) is more stringent than −v(−x) ≥ v(x)
for all x, the condition originally put forward in Kahneman and Tversky
(1979). Thinking in terms of deformation of the cumulative graph, loss
aversion guarantees that the stretching of the horizontal axis at x < 0 is at
least as large as the stretching of the horizontal axis at |x| > 0. This allows
us to use “+” segments in the losses domain where G(x) > F (x) to
counteract “−” segments in the gains domain where G(|x|) < F (|x|).
Loss aversion entails comparisons between the negative and the
positive domain. Hence, we need to constrain the sign-dependent pwfs.
Throughout this section we assume that the given values of c− , d− , c+ , and
d+ satisfy c+ < d+ and c− < d− . Under this specification, we let s be the
slope of the linear segment of w between c and d, respectively. If w is
continuous, then s = (w(d) − w(c))/(d − c) [in general
s = limε→0 [w(d − ε) − w(c + ε)]/(d − c)]. We will consider pairs of pwfs
satisfying s− ≥ s+ . That the slope of w− is larger than the slope of w+
ensures that zones of positive FSD in the negative domain can counteract
zones of negative FSD in the positive domain. The condition s− ≥ s+ is
consistent with the empirical finding that w− exhibits less deformation
than w+ (Tversky and Kahneman [1992] and Abdellaoui [2000]). For
+
the following definition, set A = B =[0, b] and calculate x+
c and xd
according to definition 4.
Definition 9. F dominates G according to P L-SD, denoted by F P L-SD G, if and
only if
+
G (x) ≥ F (x) for a ≤ x < x+
d and xc ≤ x < b; and
(2.15)
G (−x) − F (−x) ≥ F (x) − G(x), for x ≥ 0 and F (x), G(x) continuous.
(2.16)
16
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
−
Proposition 10. F P L-SD G if and only if VF ≥ VG for all v ∈ VL , w− ∈ Wcd− ,
+
and w+ ∈ Wcd+ such that s− ≥ s+ .
Graphic 2.4 (left) shows an example satisfying the conditions of the
+
proposition3 . Notice that if d+ ≤ c+ , then x+
c ≤ xd and (2.15) would
reduce to FSD.
We now take up the more challenging task of defining a condition of
stochastic dominance that combines curvature conditions together with
loss aversion. For our purposes, it is sufficient to consider the class of
prospect value functions which exhibit also loss aversion, or
VP L = VP ∩ VL . An S-shaped value function provides the possibility to use
“+” segments of small gains (small losses) to counteract “−” segments of
large gains (large losses). We call this gains-gains (losses-losses)
cancelation. Loss aversion brings the possibility of losses-gains cancelation,
i.e., utilize “+” segments of losses to cancel “−” segments of appropriately
sized gains. To know which parts of positive FSD segments in the losses
domain are available to counteract negative FSD parts in the positive
domain, we rely again on the function h(x). Recall that the segments
where h(x) is increasing are precisely those where F FSD G, so that they
are available to counteract segments in which h is decreasing. Decreasing
values of h(x) in the negative domain can only be made up for with
increasing values of h(x̂) for an x̂ between x and zero; and this is most
efficiently done for an x̂ as close to x as possible. All those pairs (x, x̂) that
can perform the losses-losses cancelation in this way are given by the
following set:
E − = {x < 0 : h(x) ≤ h(x̂), for some x̂, a ≤ x̂ < x}
Going from negative to positive, points in E − are those for which the
function h falls below a level that was reached before. Now consider a
point x < 0 not in E − . Then, h(x) increases beyond any past value of x,
implying that h0 (x) ≥ 0. Hence, points such as x are available for
losses-gains cancelation. However, according to (L), for this cancelation to
be successful it needs to be applied to a point in the positive domain with
value |x| = −x or larger. This will be the case if h(−x) ≥ h(x). We are now
ready for the full CPT SD definition, which incorporates a slightly
modified version of P W -SD, together with this last condition.
Definition 11. F dominates G according to P W L-SD, denoted by F P W L-SD G,
3. This example also illustrates that in discontinuity points, such as x = 1000, the inequality in
(2.16) need not be satisfied.
17
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
if and only if
h(x) ≤ h(0), for a ≤ x < 0,
h(x) ≥ h(0) for 0 ≤ x <
(2.17)
x+
d,
G (x) ≥ F (x) for a ≤ x <
x−
c
(2.18)
and x+
c
−
≤ x < b, and
h(−x) ≥ h(x) for x < 0 and x ∈
/E .
(2.19)
(2.20)
Proposition 12. F P W L-SD G if and only if VF ≥ VG for all v ∈ VP L ,
−
+
w− ∈ Wcd− , and w+ ∈ Wcd+ such that s− ≥ s+ .
Graphic 2.4 (right) shows an example of this kind of stochastic
dominance. We invite the reader to follow the arrows that identify the
different types of cancelation (losses-losses and losses-gains). In the
losses-losses cancelation, arrows have to be oriented right to left. In the
losses-gains cancelation, an arrow starting at x < 0 can be applied to |x| or
beyond, but never to values lower than |x|. In the gains-gains cancelation
(present in Task III [graphic 2.3] but not in Task XI [graphic 2.4]), arrows
have to be oriented left to right.
If F and G are all-gains prospects, then P W L-SD, P W -SD, and
SW -SD coincide; and if F and G are all-losses prospects, then P W L-SD,
P W -SD, and S ∗ W -SD coincide. Clearly, P W -SD implies P W L-SD, but the
converse is not true unless E − is the entire negative domain.
We finish this section with table 2.1, which summarizes properties of
the value function and the pwf that correspond to the six different forms of
stochastic dominance introduced in the working paper for the CPT model.
GRAPHIC 2.4:
Tasks VIII and XI. Task VIII involves cancelations based solely on loss aversion (WL). Task
XI requires in addition a convex value function for losses (PWL)
18
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
TABLE 2.1:
Summary of the different stochastic dominance conditions
and their requirements on the value function v and the pwf
w
SD Condition
Value Function v
pwf w–
pwf w+
(c− , 1]
P W -SD
S-shaped
Convex on
P ∗ W -SD
Inverse S-shaped
Concave on [0, d− )
[0, d− )
Convex on (c+ , 1]
Concave on [0, d+ )
Convex on (c+ , 1]
SW -SD
Concave
Concave on
S ∗ W -SD
Convex
Convex on (c− , 1]
Concave on [0, d+ )
W L-SD
Loss Aversion (L)
Inverse S-shaped
Inverse S-shaped
P W L-SD
S-shaped and (L)
Linear on (c− , d− )
Linear on (c+ , d+ )
19
3. Experimental
Applications
THE analytical SD conditions introduced above suggest experimental
designs of fine-tuned prospects through which different elements of CPT
can be tested. This section presents the experimental results of twelve
prospect comparisons or tasks that were presented to 277 individuals. The
subjects were students (undergraduates and MBA) and professionals
(business executives and lawyers). The 84 college students are from the
University of Navarra (Pamplona, Spain), majoring in economics and most
of them from Spain. The 78 MBA students are from IESE Business School
(Barcelona, Spain) representing a broad range of countries of origin. The
99 executives were participants at executive education programs at IESE.
The professional group also contains a sample of 15 lawyers who were
approached through personal contacts. The group of professionals
consists almost exclusively of Spaniards. The answers from both types of
students and the professionals are very similar so that their results are
reported together. The different tasks were divided between different
sessions so that the number of valid answers for a given task varies between
177 and 277. Subjects were not paid. According to Camerer and Hogarth
(1999), this should not significantly influence the average results for gains
in choices among risky gambles. For losses, data for a similar conclusion
are not available.
The tasks (and results) are presented in tables 3.3 through 3.6.
Prospect F always corresponds to the prospect that is preferred according
to the SD condition being tested. In the actual experiments, both the
assignment of the preferred prospect to F or G and the order of the tasks
were randomized. All tasks were printed on a questionnaire that was
handed out to the subjects. The tasks were introduced with the written
question: “Suppose that you decide to invest $10,000 either in stock F or in
stock G. Which stock would you choose, F, or G, when it is given that the
dollar gain or loss one month from now will be as follows:”. Typically,
subjects needed 20 minutes to complete the questionnaire.
The experimental part is divided into four sections, depending on
which SD conditions are applied and, most importantly, on four possible
20
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
ways in which SD tests can be interpreted. In the first section we assume
that the pwf is inverse S-shaped and the goal is to investigate the curvature
of the value function. In contrast, the second section assumes that the
value function is S-shaped and tests hypotheses on the pwf. The third
section is similar to the first, i.e., we assume that the pwf is inverse
S-shaped, but we test whether the value function exhibits loss aversion. In
the final section, we assume that all the empirical specifications of CPT for
v and w hold, including loss aversion, and interpret the results as a global
test of CPT. The hypotheses and findings of the four sections are
summarized in table 3.1.
TABLE 3.1:
Summary for each task of the joint hypotheses tested on the value function v and the pwf w
Joint Hypothesis
Not
Task
SD Condition
Value Fctn
pwf
Result
I
F SW -SD G
v ∈ VP
w+ ∈ W0.1
Not Rejected
w
Consistent with v S-shaped
G S ∗ W -SD F
v ∈ VP ∗
w+ ∈ W 0.9
Rejected
w
v∈
/ VP ∗ , v ∈
/ VConvex
v ∈ VP
w−
Not Rejected
w
Consistent with v S-shaped
v ∈ VP ∗
w− ∈ W 0.9
Rejected
w
v∈
/ VP ∗ , v ∈
/ VConcave
F P W -SD G
v ∈ VP
w ∈ W1/6
Not Rejected
w
Consistent with v S-shaped
G P ∗ W -SD F
v ∈ VP ∗
w ∈ W 2/3
Rejected
w
v∈
/ VP ∗
IV
F P W -SD G
v ∈ VP
w ∈ W0.02
Not Rejected
v
Cannot reject w ∈ W0.02
V
F P W -SD G
v ∈ VP
w ∈ W0
Rejected
v
w∈
/ W0
II
F
S ∗ W -SD
G
G S ∗ W -SD F
III
VI
F W L-SD G
v ∈ VL
w∈
VII
F W L-SD G
v ∈ VL
w∈
VIII
F W L-SD G
v ∈ VL
w∈
IX
F P W L-SD G
v ∈ VP L
w∈
X
F P W L-SD G
v ∈ VP L
w∈
XI
F P W L-SD G
v ∈ VP L
w∈
XII
F P W L-SD G
v ∈ VP L
w∈
∈ W0.1
3.1.
0.5
W0.1
0.5
W0.2
0.5
W0.2
0.5
W0.1
0.7
W0.1
0.65
W0.15
0.5
W0.2
questioned
Conclusion
Rejected
w
Loss Aversion fails
Not Rejected
w
Consistent with Loss Aversion
Not Rejected
w
Consistent with Loss Aversion
Rejected/Mixed
v, w
CPT fails
Not Rejected
v, w
Consistent with CPT
Not Rejected
v, w
Consistent with CPT
Not Rejected
v, w
Consistent with CPT
Choice of c and d
It may appear as if the exact knowledge of the parameters c and d were
critical for a sound application of the SD relations. However, it is only
necessary in the design of the tasks to choose c and d such that they are
respectively big (c) or small (d) enough such that the pwf can be safely
assumed to be concave up to d and convex from c on. Between c and d, the
pwf has to be approximately linear. The higher the c and the lower the d
one chooses, the less one assumes about the pwf, but the more restrictive is
the set of lotteries one can design. Hence, in order to increase the
21
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
freedom in choosing the prospects, we want to know how low c and how
high d can be. While the parametric forms of the pwf proposed in the
literature (Tversky and Kahneman, 1992; Prelec, 1998) have no linear
segments, we can consider some range as being approximately linear if the
differences in slope between any two points are not too large. Taking the
inflection point which has the minimum slope as the reference, we
calculate the range, delimited by c and d, for which w0 (p) is at most 50%
higher. This is a conservative estimate, if we keep in mind that in
commonly employed parametric specifications the slope of w (and v) at
different points can differ by several orders of magnitude. For the
formulation for the pwf proposed by Prelec (1998), the inflection point is
at 1/e, and w0 (1/e) = γ. Table 3.2 shows an overview of the corresponding
values for c and d. Values of c greater than 0.10 and d smaller than 0.76
should be sufficient for the empirically plausible range of γ, i.e. 0.5 to 0.7.
TABLE 3.2:
Based on the pwf, w(p)=exp[-(-ln p)γ ], the values of c and d
are the end points of the interval where w0 (p) is less than or
equal to 1.5 w0 (1/e)
γ
1
0.9
0.8
0.7
0.6
0.5
3.2.
c
0.00
0.02
0.05
0.07
0.09
0.10
d
1.00
0.98
0.90
0.84
0.79
0.76
Shape of the value function
Design
The first experimental part (Tasks I through III) aims at finding the most
representative value function, assuming that the pwf is inverse S-shaped.
Task I presents an all-gains choice, Task II an all-losses choice, and
Task III uses a mixed choice (see table 3.3 and, for Task III, see also
graphic 2.3). They are designed so that F P W -SD G. In the gains domain,
P W -SD becomes the generalized second order stochastic dominance
SW -SD. In the losses domain, P W -SD agrees with the reverse condition
S ∗ W -SD. Because EF = EG , the comparison between G and F in all three
tasks exhibits G P ∗ W -SD F . Hence, we perform two tests of the curvature
22
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
Design and Results of Tasks I, II and III
TABLE 3.3:
F
TASK I
F SW -SD G, c = 0.1
G S ∗ W -SD F , d = 0.9
TASK II
F
S∗W
-SD G, c = 0.1
G
Choice
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
0
50%
277
74
26
0
10%
1000
40%
2000
40%
3000
10%
3000
50%
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-3000
50%
-3000
10%
273
65
35
-2000
40%
-1000
40%
G SW -SD F , d = 0.9
0
50%
0
10%
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
F P W -SD G, c = 1/6
-6000
1/3
-6000
1/6
276
76
24
G P ∗ W -SD F , d = 2/3
3000
1/2
-3000
1/3
4500
1/6
4500
1/2
TASK III
of v in each task. Here, we don’t question the shape of the pwf, which is
assumed to be in Wcd , for values of c equal to 0.1 (Task I and II) and 1/6
(Task III); and of d equal to 0.9 (Tasks I and II) and 2/3 (Task III). Hence,
we test whether v is in VP for F P W -SD G and whether v is in VP ∗ for G
P ∗ W -SD F .
Results of Tasks I - III
The results of Tasks I-III are presented in the rightmost columns of
table 3.3. For each task, the number of answers is provided as well as the
percentage of individuals that chose the respective prospect. In all three
tasks, prospect F is preferred over prospect G. For the joint hypotheses on
the value function and the pwf this means the following: In Task I, G is not
preferred and, as we assume that the pwf is concave in [0, 0.9), it follows
that the most representative value function is not convex for gains. Hence,
v cannot be in VP ∗ or in VConvex . Similarly, in Task II, the hypothesis that
the value function is concave for losses is rejected, concluding that v is
neither in VP ∗ nor in VConcave . The result of Task III completes the picture
for a mixed lottery. Under the assumption that the pwf is concave in
[0, 2/3), and given that G is not preferred, the value function cannot be in
VP ∗ .
In all the three tasks, the majority of individuals prefer the P W -SD
dominating prospect F . Given the rejected shapes of the value function,
23
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
the S-shaped value function remains as a specification that is consistent
with the results. It also establishes risk aversion for gains and risk seeking
for losses as a phenomenon independent of the certainty effect.
Certainly, d = 0.9 in Tasks I and II is higher than the proposed value
of 0.76 in subsection 3.1.. However, in Tasks I and II, the nonlinear pwf
has a minor effect because the vertical stretching due to the pwf close to
p = 0 magnifies a zone of negative FSD, whereas the vertical stretching
close to p = 1 acts over a zone of positive FSD. In any case, the results of
Task III, done under d = 2/3, are sufficient to reject that the most
representative v is in VP ∗ , VConcave , or VConvex .
3.3.
Shape of the pwf
Design
The second experimental part focusses on the pwf. In the joint hypotheses
on the value function and the pwf that our SD relations test, we assume
that the value function follows the specifications of CPT, i.e. v ∈ VP , and
use P W -SD to test the convexity of the pwf. Table 3.4 shows the design of
the two corresponding tasks. Again, in both tasks we have F P W -SD G.
Task IV tests the convexity of w− on [0.02, 0.5) and of w+ on [0.02, 0.74);
and Task V tests the convexity of w− on [0, 0.5) and of w+ on [0, 0.75).
TABLE 3.4:
Design and Results of Tasks IV and V
F
TASK IV
Gain/Loss
G
Prob.
Gain/Loss
Choice
Prob.
N
F[%]
G[%]
273
81
19
F P W -SD G, c = 0.02
-6000
26%
-6000
2%
G P ∗ W -SD F , d = 0.74
3000
72%
-3000
48%
4500
2%
4500
50%
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-6000
1/4
-3000
1/2
271
37
63
3000
3/4
4500
1/2
TASK V
F P W -SD G, c = 0
G
P ∗ W -SD
F, d = 1
Results of Tasks IV and V
The results of Tasks IV and V are presented in the rightmost columns of
table 3.4 (for Task V, see also graphic 2.1). In Task IV, F is preferred by
most subjects. Hence, the hypothesis that the pwf is convex form 0.02 on
cannot be rejected. This is remarkable for the low value of c and
underscores the earlier considerations of the right choice of c in the
experimental design. In Task V, however, the majority of subjects prefers
24
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
prospect G. Thus rejecting that both w− and w+ are convex in the
corresponding intervals. Together with the results of Task IV, the
reasonable conclusion is that the pwf is not convex near the origin. Notice
that Task IV is a slight modification of Task V. In Task IV the extreme
outcomes are set equal; in F the maximum outcome has been added and
in G the minimum outcome has been added, both with a probability of
2%. This change is sufficient to reverse the preference of the majority of
subjects. That a common range is sufficient to account for most of the
effect of the pwf suggests that decision makers use the range of outcomes as a
decision criterion.
Task V is a replication of Experiment 1.III from Levy and Levy
(2002a). They use this task to perform a test on the shape of the value
function, assuming that the pwf does not produce any significant
distortion (i.e., without questioning the shape of the pwf, as we do in Tasks
I-III). Hence, they conclude that v is not in VP and, instead, endorse the
VP ∗ class of value functions. The mistake, of course, is that they implicitly
assume that the pwf is convex (or linear) around zero. Wakker (2003) and
Baucells and Heukamp (2004) show that the experimental results
presented by Levy and Levy (2002a) are compatible with CPT. But a
stronger case for the VP class, and a rejection of the VP ∗ class, is made in
Tasks III and IV, which use the SD conditions suited to test the shape of
the value function under the assumption that the pwf is inverse S-shaped.
Notice that in Tasks IV and V, prospect G dominates F according to
∗
P W -SD. This allow us to perform the following test: assume that v is in
VP ∗ , and test the concavity of the pwf. Thus, the preference for F in Task
IV could be interpreted as rejecting the hypothesis that w− is concave on
[0.02, 0.5) and w+ concave on [0.02, 0.74). However, having just rejected
the hypothesis that v is in VP ∗ , we don’t embrace this interpretation.
3.4.
Loss aversion
Design
The third experimental part consists of three tasks (VI to VIII) that focus
on loss aversion. We assume that the pwf is inverse S-shaped and test
whether the value function exhibits loss aversion. The experimental
assessment of loss aversion is a complex issue that has not received a lot of
attention in the past (see for example Schmidt and Traub, 2002) Here, we
test loss aversion employing W L-SD, introduced in proposition 10.
In Tasks VI, VII and VIII prospect F dominates G according to
W L-SD (see table 3.5 and, for Task VIII, see also graphic 2.4). The
25
MANEL BAUCELLS ALIBÉS
TABLE 3.5:
and
FRANZ H. HEUKAMP
Design and Results of Tasks VI through VIII
F
TASK VI
F W L -SD G
c = 0.1, d = 0.5
TASK VII
F W L -SD G
c = 0.2, d = 0.5
TASK VIII
Gain/Loss
G
Choice
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-1000
10%
-1000
50%
216
43
57
0
80%
1000
10%
1000
50%
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-3000
20%
-3000
50%
208
61
39
0
60%
3000
20%
3000
50%
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
F W L -SD G
-3000
20%
-3000
50%
208
64
36
c = 0.2, d = 0.5
-1000
30%
1000
30%
3000
20%
3000
50%
preference for prospect F is based therefore solely on loss aversion if one
assumes that the pwf is inverse S-shaped. While Tasks VI and VII include a
zero payoff and test loss aversion close to the origin of the value function,
Task VIII tests loss aversion beyond ±$1000. We invite the reader to draw
the cumulative distributions of F and G and check the cancelation of the
“−” areas b “+” areas.
Results of Tasks VI to VIII
The results of the tasks are presented in the rightmost columns of
table 3.5. In Task VI, the majority of subjects prefers G and the hypothesis
of a value function that incorporates loss aversion is rejected. In Tasks VII
and VIII the results are consistent with loss aversion because the majority
of subjects prefers F . One possible explanation for these results is that the
assumption about the pwfs (c = 0.1 and d = 0.5) in Tasks VI is stronger
than in Tasks VII and VIII (c = 0.2 and d = 0.5). However, c = 0.1 is
empirically plausible. A more plausible explanation is that subjects might
be following the heuristic of choosing the prospect with the highest
probability of strictly positive gains, according to which G is more
appealing in Task VI. Such an explanation would be in line with work by
Payne (2005), who finds data that contradicts a parametric specification of
CPT, but is strongly consistent with this heuristic. Still, the results of Task
VII are at odds with this heuristic. The stakes in Task VII are three times
larger than the stakes in Task VI. Thus, we can hypothesize that the degree
26
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
of loss aversion increases with stakes. There is some additional evidence
supporting this hypothesis. We can safely assume that students and
professionals differ in their assessment of the stakes so that the same dollar
amount is subjectively perceived as larger by MBAs and undergrads than by
professionals. In Task VI, MBAs and undergrads prefer G by a slight
margin, 52%, whereas the proportion of professionals violating loss
aversion is 65%, which is statistically different from 50%. A similar pattern
is observed in Task VII: MBAs and undergrads show strong loss aversion
and professionals prefer F by only a slight margin. Hence, the heuristic of
considering the probability of positive gains may apply only/specially in
the presence of small stakes (Payne, 2005, uses hypothetical choices
among graduate and undergraduate students, with outcomes varying
between -$90 and +$138).
3.5.
CPT
Design
In the fourth experimental part we assume that all the empirical
specifications of CPT for the value function and the pwf, including loss
aversion hold. We design Tasks IX to XII (see table 3.6) with prospect F
dominating G according to P W L-SD. Thus we interpret the tasks as a
global test of CPT. Specifically, in Task IX Prospect F is preferred if
besides exhibiting loss aversion, a decision maker also has a value function
that is concave for gains. Only in that case does the positive FSD zone in
the losses range cancel the negative zone in the gains range associated with
the highest outcome and do the remaining zone in the gains range cancel
each other (again, we encourage the reader to draw the cumulative
distributions of F and G and check the cancelation of the “−” areas by “+”
areas). Tasks X and XI necessitate a value function which is convex for
losses to make prospect F be preferred over prospect G. Tasks X and XI
require a sophisticated accounting between the extreme positive and
negative zones and the remaining negative zones (for Task XI, see
graphic 2.4). Finally, Task XII shows an example where the losses-gains
cancelation could be done in several ways.
Results of Tasks IX to XII
The results of the tasks are presented in the rightmost columns of
table 3.6. In Task IX, the majority of subjects prefers G although not at a
statistically significant level (The overall 46-54 response is not statistically
different from 50 percent by a one-tailed binomial test at α = 0.05). That
27
MANEL BAUCELLS ALIBÉS
TABLE 3.6:
and
FRANZ H. HEUKAMP
Design and Results of Tasks IX through XII
F
TASK IX
F P W L-SD G
c = 0.1, d = 0.5
TASK X
F P W L-SD G
G
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-500
0
1500
2000
10%
40%
40%
10%
-500
500
1000
2000
30%
20%
20%
30%
209
46
54
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-2000
30%
-2000
-1000
10%
60%
216
70
30
0
1000
60%
10%
1000
30%
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-5000
-3000
0
3000
5000
15%
30%
20%
20%
15%
-5000
-1000
35%
30%
209
74
26
5000
35%
Gain/Loss
Prob.
Gain/Loss
Prob.
N
F[%]
G[%]
-1500
1500
4500
20%
60%
20%
-1500
50%
177
77
23
4500
50%
c = 0.1, d = 0.7
TASK XI
F P W L-SD G
c = 0.15, d = 0.65
TASK XII
F P W L-SD G
c = 0.2, d = 0.5
Choice
Gain/Loss
subjects are not following CPT in this task might be explained by G having
a higher probability of strictly positive gains (Payne, 2005). In the other
three tasks, X to XII, F is clearly preferred by most of the subjects, in line
with the predictions of CPT.
For the seven tasks that show W L-SD or P W L-SD (Tasks VI through
XII) we see strong support for CPT on an individual basis: 24% of the
individuals follow CPT in all tasks and 76% follow CPT in half or more of
their choices.
28
4. Conclusions and
Extensions
THE SD conditions presented in this working paper can be used to test
joint hypotheses on the shape of the value function and the pwf (see
table 2.1). Making some assumptions on the shape of the pwf one can
easily find a simple test for the shape of the value function and viceversa.
Thus we supply a potentially powerful framework to suggest experimental
designs that isolate certain qualitative features of either the value function
or the pwf without having to estimate these functions. This adds a method
to the toolbox of an experimentalist who many times does not need to
estimate the entire value function or pwf; rather she is interested only in
falsifying a specific hypothesis about the shape of either value function or
pwf. Clearly, a different method is needed if the entire value function and
pwf needs to be explored (Abdellaoui, Bleichrodt and Paraschiv, 2004).
Three ways of testing joint hypotheses based on the SD conditions have
been presented in the second part of the working paper (see table 3.1).
a) Assuming that the empirical specifications of CPT for the pwf hold, one
can test the curvature of the value function and the presence of loss
aversion. Examples for the test of curvature are Tasks I through III and
for loss aversion Tasks VI through VIII.
b) Assuming that the value function follows the empirical specifications of
CPT, one can test the curvature of the pwf. This is what we have done
in Tasks IV and V.
c) Assuming that all the empirical specifications of CPT hold, violations of
SD actually imply violations of CPT. This way of applying the SD
conditions can be very useful to pit CPT against heuristics. This can be
the interpretation of Task VI and IX, in which the violation of SD
actually suggests that, in those tasks, individuals resort to ad–hoc
heuristics.
A restrictive feature of our method is that the designs of the tasks
which use the SD notions always imply adding common extreme
29
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
outcomes. This requires prospects with at least three outcomes, which
precludes the use of our SD conditions to test CPT in simpler prospects. In
addition, working with the parameters c and d is in itself a drawback
compared to other methods of measuring weighting functions
(Abdellaoui, 2000). This is rooted in the nature of the extension of the
SD-techniques to CPT which require joint curvature conditions that are at
odds with the double curvature observed for the pwf. Even so, the
parameters c and d do not have to be known exactly but can be chosen
from a quite large interval. Furthermore, in our presentation, we focus on
a family of pwfs that includes the inverse S-shaped pwf, but that also
contains empirically implausible forms. The pwf class could be tightened,
e.g., by restricting W to inverse S-shaped pwfs that cross the 45-degree line
from above. We leave for future exploration the SD conditions that
characterize such preferences.
Extending our results to the rank-dependent framework is
straightforward: Chew, Karni, and Safra (1987) have shown that S-SD
applies to rank-dependent models for concave w = w− .
The experimental results in this paper confirm that it is important to
account for the pwf, especially in choices where the lower and upper
distortions of the pwf tend to reinforce the same prospect (see for example
Task IV). In those cases, very small probability changes in the extreme
outcomes (here 2%) suffice to reverse preferences of individuals. SD
conditions are very adequate to uncover violations of a choice theory,
which is a very active area of research (see Wu and
Markle, 2004; Birnbaum, In press; Payne, 2005). For example, while we
confirm the existence of loss aversion (which clearly shows if the
probability of having positive outcomes is similar or the same in both
choices), we document violations of loss aversion attributable to subjects
using the probability of strictly positive gains as a criteria. This last effect
shows specially in tasks with relatively small stakes.
30
Appendix:
Proofs
Proof of proposition 1. (⇒) Recall that h0 (x) = G(x) − F (x) and
v 0 (x) ≥ 0, so that, in view of (2.1), the goal is to use (2.3) and (2.4) to show
that the integration in the range where h0 (x) < 0 is compensated by the
integration in the range where h0 (x) ≥ 0, i.e.,
∆{x:h0 (x)≥0} + ∆{x:h0 (x)<0} ≥ 0. We now define a cancelation function H,
that to each point in {x : h0 (x) < 0} will associate a point H(x) having
h0 (H(x)) ≥ 0. If x ∈ A, then let H(x) = inf{x0 > x : h(x0 ) = h(x)}; and if
x ∈ B, then let H(x) = sup{x0 < x : h(x0 ) = h(x)}. By continuity of h,
right-continuity of h0 , (2.3), and (2.4), H is a well defined, one-to-one
function with points in A (resp. B) having their image in A (resp. B). We
now claim that in the domain of H, v 0 (H(x)) ≥ v 0 (x). If x ∈ A, then
H(x) > x, which together with the convexity of v in A, proves the claim;
and if x ∈ B, then H(x) < x, which together with the concavity of v in B,
proves the claim.
Let D◦ be the domain of H after having excluded the following
exceptional points: a, 0, b, discontinuity points of F or G, and points x
where h0 (H (x)) = 0. A moment’s reflection reveals that if a point is not
exceptional, then it is in the interior of D◦ , i.e., D◦ is open, and that the
integration of ∆[a,b] in exceptional points is zero (they are isolated point of
measure zero). Moreover, if x ∈ D◦ , then H(x), h0 (x), and h0 (H(x)) are
continuous, and h0 (H (x)) > 0. Using h(x) = h(H(x)), we have that
h0 (x) = h0 (H(x))H 0 (x). Hence H 0 (x) = h0 (x)/h0 (H (x)) is well defined in
D◦ . We now write
Z
Z
Z
0
0
0
h (x)dx =
h (H(x))|H (x)|dx = −
h0 (x)dx
(4.1)
H(D◦ )
D◦
D◦
where the first equality follows from the change of variable formula for
Lebesgue integrals [Strichartz (1995, p.719)], and the second from
h0 (x) = −h0 (H (x)) |H 0 (x)|. Using (4.1) and the claim, we conclude that
∆H(D◦ ) + ∆D◦ ≥ 0.
31
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
(⇐) Assume (2.3) fails. Hence, h(x1 ) > h(a1 ) for some x1 ∈ A. Let
v(x) = min{0, max{x − a1 , x1 − a1 }}
(4.2)
which is convex in A. Applying (2.1) we find ∆[a,b] = h(a1 ) − h(x1 ) < 0, a
contradiction.
Finally, assume (2.4) fails. Hence, h(x0 ) < h(b0 ) for some x0 ∈ B. Let
v(x) = max{0, min{x − b0 , x0 − b0 }}
(4.3)
which is concave in B. Thus, ∆[a,b] = h(x0 ) − h(b0 ) < 0.
Proof of proposition 3. Each of the four results is a particular case
of proposition 1, after defining the intervals A and B. For example,
conditions (2.3) and (2.4), as applied to A = [a, 0] and B = [0, b], agree
with the definition of P -SD. Hence VF ≥ VG for all v convex in [a, 0] and
concave in [0, b], which is precisely the VP class. To obtain P ∗ -SD, S-SD,
and S ∗ -SD, use A = [0, b] and B = [a, 0]; A = ∅ and B = [a, b]; and
A = [a, b] and B = ∅, respectively.
−
Proof of proposition 5. The definition of x−
c ensures that w (p) is
convex (concave) in points where p is equal to either F (x) or G(x),
−
−
−
x−
c < x < 0 (a ≤ x < xd ). Hence, [w (G(x)) − w (F (x))]/[G(x) − F (x)]
0
is non-decreasing (non-increasing). Of course, v (x) ≥ 0 is non-decreasing
(non-increasing) in A (B), and the first part of the result follows.
−
Similarly, the definition of x+
c ensures that w (p) is concave (convex) in
+
points where p is equal to either F (x) or G(x), 0 ≤ x < x+
c (xd < x < b).
Hence, [w− (G(x)) − w− (F (x))]/[G(x) − F (x)] is non-increasing
(non-decreasing). Of course, v 0 (x) ≥ 0 is non-increasing (non-decreasing)
in B (A), and the second part of the result follows.
Proof of proposition 6. (⇒) We take the function h and repeat the
construction of H as in the proof of proposition 1. By (2.13) and (2.14),
points in (a, 0) and (0, b) have their image in [a, 0] and [0, b], respectively.
Hence, we can let D◦− = D◦ ∩ (a, 0) and D◦+ = D◦ ∩ (0, b) and, using a
change of variable, write
Z
Z
0
h (x)dx = −
h0 (x)dx, and
(4.4)
◦−
◦−
H(D )
D
Z
Z
0
h (x)dx = −
h0 (x)dx.
(4.5)
H(D◦+ )
D◦+
We now wish to combine proposition 5 with (4.4) and (4.5) to produce
w
w
w
∆w
H(D◦− ) + ∆D◦− ≥ 0 and ∆H(D◦+ ) + ∆D◦+ ≥ 0. To do so, we let x̂ = H(x)
and check the following. If x ∈ A ∩ D◦− , then x̂ > x and, from (2.11),
◦−
both x, x̂ ∈ A ∩ [x−
c , 0]. If x ∈ B ∩ D , then x̂ < x and both
32
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
◦+
x, x̂ ∈ B ∩ [a, x−
d ]. If x ∈ A ∩ D , then x̂ > x and, from (2.11), both
◦+
x, x̂ ∈ A ∩ [x+
d , b]. Finally, if x ∈ B ∩ D , then x̂ < x and both
+
x, x̂ ∈ B ∩ [0, xc ]. Hence, the desired monotonicity property holds.
(⇐) The proof of proposition 1 shows that both (2.3) and (2.4) are
necessary. Assume (2.11) fails. Hence, h0 (x1 ) < 0 for a0 ≤ x1 < x−
c or
0 (x), there is an ε > 0 such that
â0 ≤ x1 < x+
.
By
right-continuity
of
h
d
x1 + ε < a1 and h0 (x) < 0 for x1 ≤ x < x1 + ε. Because â0 ≥ 0, both x1
and x1 + ε have the same sign. If x1 < 0, then let pc = min{F (x1 + ε), c− }
and pd = 1; and if x1 ≥ 0, then let pc = 0 and
1 − pd = min{F (x1 + ε), 1 − d+ }. Let v be as in (4.2), w− as
w(p) = min{1, p/pc }
(4.6)
with the understanding that if pc = 0, then w(p) = 1, 0 < p ≤ 1; and w+ as
w(p) = max{0, (p − pd )/(1 − pd )}
(4.7)
with the understanding that if pd = 1, then w(p) = 0, 0 ≤ p < 1. We check
that v is convex in A, w− convex in (c− , 1] and concave throughout, and
w+ concave in [0, d+ ) and convex throughout. v ensures that only
integration in [x1 , a1 ] matters. w ensures that points x > x1 + ε having
h0 (x) > 0 do not yield a positive contribution because if x < 0, then w− is
constant whenever G(x) > F (x1 + ε) ≥ pc ; and if x ≥ 0, then w+ is
constant whenever G(x) > F (x1 + ε) ≥ 1 − pd . That h0 (x) < 0 in
w
[x1 , x1 + ε] implies ∆w
[a,b] ≤ ∆[x1 ,x1 +ε] < 0, a contradiction.
+
If (2.12) fails, then h0 (x1 ) < 0 for x−
d ≤ x1 < b̂1 or xc ≤ x1 < b1 . As before,
there is an ε > 0 such that x1 + ε < b1 and h0 (x) < 0 for x1 ≤ x < x1 + ε.
Let x0 = x1 + ε. Because b̂1 ≤ 0, both x1 and x0 have the same sign. If
x1 < 0, then let pd = max{G(x1 ), d− } and pc = 0; and if x1 ≥ 0, then let
pd = 1 and 1 − pc = max{G(x1 ), 1 − c+ }. Let v, w− , and w+ be as in (4.3),
(4.7), and (4.6), respectively. Notice that v is concave in B, w− concave in
[0, d− ) and convex throughout, and w+ convex in (c+ , 1] and concave
throughout. v ensures that only integration in [b0 , x0 ] matters. w ensures
that points x < x1 having h0 (x) > 0 do not yield a positive contribution
because if x < 0, then w− is constant whenever pd ≥ G(x1 ) > F (x); and if
x ≥ 0, then w+ is constant whenever 1 − pc ≥ G(x1 ) > F (x). That
w
h0 (x) < 0 in [x0 − ε, x0 ] implies ∆w
[a,b] ≤ ∆[x0 −ε,x0 ] < 0, a contradiction.
If (2.13) fails, then h(x1 ) > h(0) for a0 ≤ x1 < 0. Let v be as in (4.2),
which is convex in A, w− (p) = p, and w+ (p) = 0, 0 ≤ p < 1, which are
linear throughout. (2.5) produces ∆w
[a,b] = h(0) − h(x1 ) < 0, a
contradiction. Finally, if (2.14) fails, then h(x0 ) < h(0) for some x0 ,
33
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
0 ≤ x0 < b1 . Let v be as in (4.3), which is concave in B, w− (p) = 0,
0 ≤ p < 1, and w+ (p) = p, which are linear throughout. (2.5) yields
∆w
[a,b] = h(x0 ) − h(0) < 0, a contradiction.
Proof of proposition 8. Each of the four results is a particular case
of proposition 6, after defining the intervals A and B. For example, letting
A = [a, 0] and B = [0, b], we obtain the conditions (2.3), (2.4) of P -SD, and
+
(2.11) implies that h0 (x) ≤ 0 for a0 ≤ x < x−
c and xc ≤ x < b1 , which
yields P W -SD. Notice that conditions (2.13) and (2.14) do not apply to
P -SD and P ∗ -SD. To obtain P ∗ -SD, S-SD, and S ∗ -SD, use A = [0, b] and
B = [a, 0]; A = ∅ and B = [a, b]; and A = [a, b] and B = ∅, respectively.
Proof of proposition 10. (⇒) If w ∈ Wcd , c < d, and
1 > G(x) ≥ F (x) > 0, then
w− (G(x)) − w− (F (x))
G(x) − F (x)
+
w (1 − F (x)) − w+ (1 − G(x))
G(x) − F (x)
≥ s− , a ≤ x < 0; and
(4.8)
≥ s+ , 0 ≤ x < b,
(4.9)
−
+
+
with equality if x+
d ≤ x < xc . This property of w, together with s ≥ s ,
0
0
v (x) ≥ v (−x) for x < 0, and (2.16), produces
∆w
[a,b]
Z
0
≥
−
0
b
Z
s [G(x) − F (x)]v (x)dx +
a
Z
s+ [G(x) − F (x)]v 0 (x)dx
0
0
≥
−
0
Z
s [G(x) − F (x)]v (−x)dy +
a
Z
≥
b
s+ [G(x) − F (x)]v 0 (x)dx
0
b
s+ [G(−x) − F (−x) + G(x) − F (x)]v 0 (x)dx ≥ 0.
0
(⇐) Condition (2.15) can fail in the following three independent
ways. First, assume G(x0 ) < F (x0 ) for some x0 , a ≤ x0 < 0. By
right-continuity of h0 (x), there is an ε > 0 such that x0 + ε < 0 and
h0 (x) < 0 for x0 ≤ x < x0 + ε. Let x1 = x0 + ε and consider
v(x) = min{0, max{x − x1 , x0 − x1 }} which is in VL ; and w(p) = p which
w
are linear throughout. This produces ∆w
[a,b] = ∆[x0 ,x1 ] < 0.
Second, assume F (x0 ) > G(x0 ) for some x0 , 0 ≤ x < x+
d . There exists ε
such that h0 (x) < 0 for x0 ≤ x < x0 + ε. Let
1 − pd = min{F (x0 + ε), 1 − d+ } and consider v(x) = x − x0 if x ≥ x0 ;
v(x) = 0 if −x0 ≤ x < x0 ; and v(x) = x + x0 if x < −x0 ; w− (p) = 0,
0 ≤ p < 1; and w+ as in (4.7). v is in VL , w− is concave in [0, 1) and convex
throughout, and w+ is concave in [0, d+ ) and convex throughout.
34
STOCHASTIC DOMINANCE AND CUMULATIVE PROSPECT THEORY
w
Moreover, s− = s+ = 0. By construction, ∆w
[a,b] ≤ ∆[x0 ,x0 +ε] < 0, which
yields the desired contradiction.
Third, assume F (x0 ) > G(x0 ), x+
c ≤ x0 < b. There exists ε such that
0
h (x) < 0 for x0 ≤ x < x0 + ε. Let x1 = x0 + ε and consider
v(x) = max{−x1 , min{x, x1 }}. Let , w− (p) = 0, 0 ≤ p < 1. Finally, let
1 − pc = max{G(x0 ), 1 − c+ } and w+ as in (4.6). v is in VL , w− is concave
in [0, 1) and convex throughout, and w+ convex in (c+ , 1] and concave
w
throughout. Moreover, s− = s+ = 0. By construction, ∆w
[a,b] ≤ ∆[x0 ,x1 ] < 0,
a contradiction.
Finally, assume (2.15) holds but (2.16) fails. Then,
+
0 ≤ G(−x0 ) − F (−x0 ) < F (x0 ) − G(x0 ) for some x0 , x+
d ≤ x0 < xc .
Notice that x0 > 0 and that F and G are continuous at x0 . Hence, there is
an ε > 0 such that G(−x) − F (−x) < F (x) − G(x) for
0 < x0 − ε ≤ x0 ≤ x0 + ε]. Define v(x) ∈ VL using continuity, v(0) = 0,
v 0 (x) = 1 for x ∈ [x0 , x0 +R ε] ∪ [−x0 − ε, −x0 ], and
otherwise; and
R x00 +ε
−x0
[G(t)
−
F
(t)]dt
+
[G(t) − F (t)]dt < 0.
w(p) = p. Then, ∆w
=
[a,b]
−x0 −ε
x0
Proof of proposition 12. (⇒) First let
E + = {x ≥ 0 : h(x) ≥ h(x̂) for some x̂, 0 ≤ x̂ < x},
which will be the set where the gains-gains cancelation takes place (the
losses-losses cancelation takes place in E − ). We define the cancelation
function HL (x), whose domain are the points having h0 (x) < 0, as follows.
If x ∈ [a, 0], then let HL (x) = inf{x0 > x : h(x0 ) = h(x)}; if x ∈ E + , let
HL (x) = sup{x0 < x : h(x0 ) = h(x)}; and if x ∈ [0, b]\E + , then let
HL (x) = sup{x0 ∈ [a, 0]\E − : h(x0 ) = h(x)}. As argued in the proof of
proposition 1, HL0 (x) is well defined on D◦ , an open dense set of
{x : h0 (x) < 0}. Let x̂ = HL (x). We now claim that v 0 (x̂) ≥ v 0 (x), x ∈ D◦ .
By (2.17), if x ∈ [a, 0], then x < x̂ ∈ E − ; and if x ∈ E + , then x > x̂ ∈ E + .
Because v ∈ VP , the claim holds in [a, 0] ∪ E + . Finally, if x ∈ [0, b]\E + ,
then x̂ ∈ [a, 0]\E − and, by (2.20), −x ≤ x̂ ≤ 0. Because v ∈ VL , the claim
holds in [0, b]\E + . If x̂ < 0 ≤ x, then notice that, by (2.18) and (2.19),
+
x+
d ≤ x < xc . Hence, from (4.8) and (4.9),
w− (G(x̂)) − w− (F (x̂))
w+ (1 − F (x)) − w+ (1 − G(x))
≥ s− ≥ s+ =
.
G(x̂) − F (x̂)
G(x) − F (x)
(4.10)
Using the change of variable formula, v 0 (x̂) ≥ v 0 (x), and (4.10) produces
w
∆w
HL (D◦ ) ≥ ∆D◦ .
(⇐) The proof of proposition 1 shows that (2.17) is necessary; and
the proofs of propositions 6 and 10 show that (2.19) is necessary for
35
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
+
a ≤ x < x−
c and xc ≤ x < b, respectively. If condition (2.18) fails, then
h(x0 ) < h(0) for some x0 , 0 ≤ x0 < x+
d . By continuity, x0 > 0, and by
differentiability, one can assume that h0 (x0 ) < 0, so that F (x0 ) > 0. Let
v(x) = max{−x0 , min{x, x0 }}, w− (p) = 0, 0 ≤ p < 1, and w+ as in (4.7)
using 1 − pd = min{F (x0 ), 1 − d+ }. We check that v ∈ VP L , w− is linear,
and w+ is concave in [0, d− ) and convex throughout, and that
s− = s+ = 0. If d+ < 1, then pd < 1 and, by construction,
∆w
[h(x0 ) − h(0)]/(1 − pd ) < 0; and if d+ = 1, then
[0,x0 ] = R
x0
∆w
[0,x0 ] = 0 −F (x)dx < 0, a contradiction. Finally, assume (2.20) fails.
Then, h(−x1 ) < h(x1 ) for some x1 < 0 and x1 ∈
/ E − . Consider
v(x) = max{x1 , min{x, −x1 }} ∈ VP L and w(p) = p. By construction,
∆w
[a,b] = h(−x1 ) − h(x1 ) < 0, a contradiction.
36
References
ABDELLAOUI, M. (2000): “Parameter-free Elicitation of Utility and Probability Weighting
Functions”, Management Science, 46(11), 1497–1512.
H. BLEICHRODT and C. PARASCHIV (2004): “Measuring Loss Aversion: A
Parameter-free Approach”, Working Paper.
F. VOSSMANN and M. WEBER (In Press): “Choice-Based Elicitation and Decomposition
of Decision Weights for Gains and Losses under Uncertainty”, Management Science.
BAUCELLS, M. and F. HEUKAMP (2004): “Reevaluation of the Results of Levy and Levy
(2002a)”, Organizational Behavior and Human Decision Processes, 94(1), 15–21.
BENARTZI, S. and R. THALER (1995): “Myopic Loss Aversion and the Equity Premium
Puzzle”, Quarterly Journal of Economics, 110(1), 75–92.
BIRNBAUM, M. (In press): “Three New Tests of Independence that Differentiate Models of
Risky Decision Making”, Management Science.
BOWMAN, D., D. MINEHART and M. RABIN (1999): “Loss Aversion in a
Consumption-Savings Model”, Journal Economic Behavior and Organization, 38(2),
155–178.
CAMERER, C. (1995): “Individual Decision Making”, in Handbook of Experimental Economics,
ed. by J. Kagel, and A. Roth, 617–703, Princeton University Press.
and R. HOGARTH (1999): “The Effects of Financial Incentives in Experiments: A
Review and Capital-Labor-Production Framework”, Journal of Risk and Uncertainty,
19(1-3), 7–42.
CHEW, S., E. KARNI and Z. SAFRA (1987): “Risk Aversion in the Theory of Expected Utility
with Rank Dependent Probabilities”, Journal of Economic Theory, 42(2), 370–381.
GONZALEZ, R. and G. WU (1999): “On the Shape of the Probability Weighting Function”,
Cognitive Psychology, 38(1), 129–166.
KAHNEMAN, D. and A. TVERSKY (1979): “Prospect Theory: An Analysis of Decision under
Risk”, Econometrica, 47(2), 263–291.
LANGER, T. and M. WEBER (2001): “Prospect Theory, Mental Accounting and Differences
in Aggregated and Segregated Evaluation of Lottery Portfolios”, Management Science,
47(5), 716–733.
LATTIMORE, P., J. BAKER and A. WITTE (1992): “The Influence of Probability on Risky
Choice - A Parametric Examination”, Journal of Economic Behavior and Organization, 17,
377–400.
LEVY, H. (1992): “Stochastic Dominance and Expected Utility: Survey and Analysis”,
Management Science, 38(4), 555–593.
and M. LEVY (2002a): “Experimental Test of the Prospect Theory Value Function: A
Stochastic Dominance Approach”, Organizational Behavior and Human Decision Processes,
89(2), 1058–1081.
and M. LEVY (2002b): “Prospect Theory: Much ado about nothing?”, Management
Science, 48(10), 1334–1349.
and Z. WIENER (1998): “Stochastic Dominance and Prospect Dominance with
Subjective Weighting Functions”, Journal of Risk and Uncertainty, 16(2), 147–163.
PAYNE, J. (2005): “It is whether you win or lose: The importance of the overall probabilities
of winning or losing in risky choice”, Journal of Risk and Uncertainty, 30, 5–19.
37
MANEL BAUCELLS ALIBÉS
and
FRANZ H. HEUKAMP
PRELEC, D. (1998): “The Probability Weighting Function”, Econometrica, 66(3), 497–528.
QUIGGIN, J. (1993): Generalized Expected Utility Theory/The Rank-Dependent Expected Utility
model, Kluwer-Nijhoff, Amsterdam.
ROTHSCHILD, M. and J. STIGLITZ (1970): “Increasing Risk. I. A Definition”, Journal of
Economic Theory, 2, 225–243.
SCHMIDT, U. and S. TRAUB (2002): “An Experimental Test of Loss Aversion”, Journal of
Risk Uncertainty, 25(3), 233–249.
STRICHARTZ, R. (1995): The Way of Analysis, Jones and Bartlett Publishers.
TVERSKY, A. and D. KAHNEMAN (1992): “Advances in Prospect Theory - Cumulative
Representation of Uncertainty”, Journal of Risk and Uncertainty, 5, 297–323.
and P. WAKKER (1995): “Risk Attitudes and Decision Weights”, Econometrica, 63(6),
1255–1280.
WAKKER, P. (2003): “The Data of Levy and Levy (2002), "Prospect Theory: Much Ado
about Nothing?" Actually Support Prospect Theory”, Management Science, 49(7),
979–981.
WAKKER, P. and A. TVERSKY (1993): “An Axiomization of Cumulative Prospect Theory”,
Journal of Risk and Uncertainty, 49(7), 147–176.
WU, G. and A. MARKLE (2004): “An Empirical Test of Gain-Loss Separability in Prospect
Theory”, Working Paper.
J. ZHANG and R. GONZÁLEZ (2004): “Decision under Risk”, in Blackwell Handbook of
Judgment and Decision Making, ed. by D. Koehler, and N. Harvey, 399–423. Blackwell,
Oxford.
38
A B O U T
MANEL BAUCELLS ALIBÉS
T H E
A U T H O R S*
is associate professor at IESE Business
School (Barcelona). His education includes a degree in mechanical engineering (Polytechnic University of Catalonia, Barcelona),
an MBA (IESE, Barcelona), and a PhD in management (University of California at Los Angeles, 1999). For two academic years
(2001-03), he held a double appointment as assistant professor at
IESE and adjunct professor at the Fuqua School of Business, Duke
University. He won the student paper competition of the Decision
Analysis Society in 2001, and mentored the winner of the competition in 2003. He is a council member of the Decision Analysis Society of INFORMS. He has taught decision analysis in MBA and executive education programs in Spain, the United States, Germany
and Nigeria. His research interests cover both normative and descriptive aspects of decision making. He is associate editor for the
journals Management Science and Operations Research.
E-mail: [email protected]
FRANZ H . HEUKAMP
received his PhD from the Massachusetts In-
stitute of Technology (MIT) and has been assistant professor of
management at IESE Business School. His research has focused
on engineering mechanics and behavioral decision making under
uncertainty. He has published articles in leading professional journals, such as Organizational Behavior and Human Decision Processes
and International Journal of Solids and Structures. He received the
2003 student paper award of the Decision Analysis Society and was
a Hugh Hampton Scholar during his PhD at MIT.
E-mail: [email protected]
Any comments on the contents of this paper can be addressed to Manel
Baucells at [email protected].
* The authors wish to thank Robin Hogarth (UPF), the Decision Sciences group at Duke, the EURO/INFORMS Istanbul conference, the
INFORMS meeting Atlanta and the BDRM 2004 conference participants
for helpful discussions; and Javier Gómez-Biscarri, Jordi Utgés, César
Beltrán de Miguel, Tony Salvia (all IESE Business School) and Javier
Vidal-Quadras for their support in conducting the experiments. Manel
Baucells is also grateful to the BBVA Foundation for financial support.
D O C U M E N T O S
D E
T R A B A J O
NÚMEROS PUBLICADOS
DT 01/02
Trampa del desempleo y educación: un análisis de las relaciones entre los efectos
desincentivadores de las prestaciones en el Estado del Bienestar y la educación
Jorge Calero Martínez y Mónica Madrigal Bajo
DT 02/02
Un instrumento de contratación externa: los vales o cheques.
Análisis teórico y evidencias empíricas
Ivan Planas Miret
DT 03/02
Financiación capitativa, articulación entre niveles asistenciales
y descentralización de las organizaciones sanitarias
Vicente Ortún Rubio y Guillem López-Casasnovas
DT 04/02
La reforma del IRPF y los determinantes de la oferta laboral
en la familia española
Santiago Álvarez García y Juan Prieto Rodríguez
DT 05/02
The Use of Correspondence Analysis in the Exploration
of Health Survey Data
Michael Greenacre
DT 01/03
¿Quiénes se beneficiaron de la reforma del IRPF de 1999?
José Manuel González-Páramo y José Félix Sanz Sanz
DT 02/03
La imagen ciudadana de la Justicia
Juan José Toharia Cortés
DT 03/03
Para medir la calidad de la Justicia (I): Abogados
Juan José García de la Cruz Herrero
DT 04/03
Para medir la calidad de la Justicia (II): Procuradores
Juan José García de la Cruz Herrero
DT 05/03
Dilación, eficiencia y costes: ¿Cómo ayudar a que la imagen de la Justicia
se corresponda mejor con la realidad?
Santos Pastor Prieto
DT 06/03
Integración vertical y contratación externa en los servicios
generales de los hospitales españoles
Jaume Puig-Junoy y Pol Pérez Sust
DT 07/03
Gasto sanitario y envejecimiento de la población en España
Namkee Ahn, Javier Alonso Meseguer y José A. Herce San Miguel
DT 01/04
Métodos de solución de problemas de asignación de recursos sanitarios
Helena Ramalhinho Dias Lourenço y Daniel Sierra de la Figuera
DT 01/05
Licensing of University Inventions: The Role of a Technology Transfer Office
Inés Macho-Stadler, David Pérez-Castrillo, Reinhilde Veugelers
DT 02/05
Estimating the Intensity of Price and Non-Price Competition in Banking:
An Application to the Spanish Case
Santiago Carbó Valverde, Juan Fernández de Guevara Radoselovics, David Humphrey,
Joaquín Maudos Villarroya
DT 03/05
Sistemas de pensiones y fecundidad: Un enfoque de generaciones solapadas
Gemma Abío Roig y Concepció Patxox Cardoner
DT 04/05
Análisis de los factores de exclusión social
Joan Subirats i Humet (Dir.), Ricard Gomà Carmona y Joaquim Brugué Torruella (Coords.)
DT 05/05
Riesgos de exclusión social en las Comunidades Autónomas
Joan Subirats i Humet (Dir.), Ricard Gomà Carmona y Joaquim Brugué Torruella (Coords.)
DT 06/05
A Dynamic Stochastic Approach to Fisheries Management Assessment:
An Application to some European Fisheries
José M. Da-Rocha Álvarez y María-José Gutiérrez Huerta
DT 07/05
The New Keynesian Monetary Model: Does it Show the Comovement
between Output and Inflation in the U.S. and the Euro Area?
Ramón María-Dolores Pedrero y Jesús Vázquez Pérez
DT 08/05
The Relationship between Risk and Expected Return in Europe
Ángel León Valle, Juan Nave Pineda y Gonzalo Rubio Irigoyen
DT 09/05
License Allocation and Performance in Telecommunications Markets
Roberto Burguet Verde
DT 10/05
Procurement with Downward Sloping Demand: More Simple Economics
Roberto Burguet Verde
DT 11/05
Technological and Physical Obsolescence and the Timing of Adoption
Ramón Caminal Echevarría
DT 01/06
El efecto de la inmigración en las oportunidades de empleo
de los trabajadores nacionales: Evidencia para España
Raquel Carrasco Perea, Juan Francisco Jimeno Serrano y Ana Carolina Ortega Masagué
DT 02/06
Inmigración y pensiones: ¿Qué sabemos?
José Ignacio Conde-Ruiz, Juan Francisco Jimeno Serrano y Guadalupe Valera Blanes
DT 03/06
A Survey Study of Factors Influencing Risk Taking Behavior
in Real World Decisions under Uncertainty
Manel Baucells Alibés y Cristina Rata
DT 04/06
Measurement of Social Capital and Growth:
An Economic Methodology
Francisco Pérez García, Lorenzo Serrano Martínez, Vicente Montesinos Santalucía
y Juan Fernández de Guevara Radoselovics
DT 05/06
The Role of ICT in the Spanish Productivity Slowdown
Matilda Mas Ivars y Javier Quesada Ibáñez
DT 06/06
Cross-Country Comparisons of Competition and Pricing Power
in European Banking
David Humphrey, Santiago Carbó Valverde, Joaquin Maudos Villarroya y Philip Molyneux
DT 07/06
The Design of Syndicates in Venture Capital
Giacinta Cestone, Josh Lerner y Lucy White
DT 08/06
Efectos de la confianza en la información contable sobre el coste de la deuda
Belén Gill de Albornoz Noguer y Manuel Illueca Muñoz
DT 09/06
Relaciones sociales y envejecimiento saludable
Ángel Otero Puime, María Victoria Zunzunegui Pastor, François Béland,
Ángel Rodríguez Laso y Marí́a Jesús García de Yébenes y Prous
DT 10/06
Ciclo económico y convergencia real en la Unión Europea:
Análisis de los PIB per cápita en la UE-15
José Luis Cendejas Bueno, Juan Luis del Hoyo Bernat, Jesús Guillermo Llorente Álvarez,
Manuel Monjas Barroso y Carlos Rivero Rodríguez
DT 11/06
Esperanza de vida en España a lo largo del siglo XX:
Las tablas de mortalidad del Instituto Nacional de Estadística
Francisco José Goerlich Gisbert y Rafael Pinilla Pallejà
DT 12/06
Convergencia y desigualdad en renta permanente y corriente: Factores determinantes
Lorenzo Serrano Martínez
DT 13/06
The Common Agricultural Policy and Farming in Protected Ecosystems:
A Policy Analysis Matrix Approach
Ernest Reig Martínez y Vicent Estruch Guitart
DT 14/06
Infrastructures and New Technologies as Sources of Spanish Economic Growth
Matilde Mas Ivars
DT 15/06
Cumulative Dominance and Heuristic Performance
in Binary Multi-Attribute Choice
Manel Baucells Alibés, Juan Antonio Carrasco López y Robin M. Hogarth
DT 16/06
Dynamic Mixed Duopoly: A Model Motivated by Linux versus Windows
Ramon Casadesus-Masanell y Pankaj Ghemawat
DT 01/07
Social Preferences, Skill Segregation and Wage Dynamics
Antonio Cabrales Gaitia, Antoni Calvó-Armengol y Nicola Pavoni
2007-02
15/1/07
10:13
Página 1
Documentos
de Trabajo
2
Documentos
de Trabajo
2
2007
Manel Baucells Alibés
Franz H. Heukamp
Stochastic Dominance
and Cumulative
Prospect Theory
Gran Vía, 12
48001 Bilbao
Tel.: 94 487 52 52
Fax: 94 424 46 21
Paseo de Recoletos, 10
28001 Madrid
Tel.: 91 374 54 00
Fax: 91 374 85 22
[email protected]
www.fbbva.es